2 * Memory Migration functionality - linux/mm/migration.c
4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
6 * Page migration was first developed in the context of the memory hotplug
7 * project. The main authors of the migration code are:
9 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10 * Hirokazu Takahashi <taka@valinux.co.jp>
11 * Dave Hansen <haveblue@us.ibm.com>
15 #include <linux/migrate.h>
16 #include <linux/module.h>
17 #include <linux/swap.h>
18 #include <linux/swapops.h>
19 #include <linux/pagemap.h>
20 #include <linux/buffer_head.h>
21 #include <linux/mm_inline.h>
22 #include <linux/nsproxy.h>
23 #include <linux/pagevec.h>
24 #include <linux/rmap.h>
25 #include <linux/topology.h>
26 #include <linux/cpu.h>
27 #include <linux/cpuset.h>
28 #include <linux/writeback.h>
29 #include <linux/mempolicy.h>
30 #include <linux/vmalloc.h>
31 #include <linux/security.h>
32 #include <linux/memcontrol.h>
33 #include <linux/syscalls.h>
37 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
40 * Isolate one page from the LRU lists. If successful put it onto
41 * the indicated list with elevated page count.
44 * -EBUSY: page not on LRU list
45 * 0: page removed from LRU list and added to the specified list.
47 int isolate_lru_page(struct page
*page
, struct list_head
*pagelist
)
52 struct zone
*zone
= page_zone(page
);
54 spin_lock_irq(&zone
->lru_lock
);
55 if (PageLRU(page
) && get_page_unless_zero(page
)) {
59 del_page_from_active_list(zone
, page
);
61 del_page_from_inactive_list(zone
, page
);
62 list_add_tail(&page
->lru
, pagelist
);
64 spin_unlock_irq(&zone
->lru_lock
);
70 * migrate_prep() needs to be called before we start compiling a list of pages
71 * to be migrated using isolate_lru_page().
73 int migrate_prep(void)
76 * Clear the LRU lists so pages can be isolated.
77 * Note that pages may be moved off the LRU after we have
78 * drained them. Those pages will fail to migrate like other
79 * pages that may be busy.
86 static inline void move_to_lru(struct page
*page
)
88 if (PageActive(page
)) {
90 * lru_cache_add_active checks that
91 * the PG_active bit is off.
93 ClearPageActive(page
);
94 lru_cache_add_active(page
);
102 * Add isolated pages on the list back to the LRU.
104 * returns the number of pages put back.
106 int putback_lru_pages(struct list_head
*l
)
112 list_for_each_entry_safe(page
, page2
, l
, lru
) {
113 list_del(&page
->lru
);
121 * Restore a potential migration pte to a working pte entry
123 static void remove_migration_pte(struct vm_area_struct
*vma
,
124 struct page
*old
, struct page
*new)
126 struct mm_struct
*mm
= vma
->vm_mm
;
133 unsigned long addr
= page_address_in_vma(new, vma
);
138 pgd
= pgd_offset(mm
, addr
);
139 if (!pgd_present(*pgd
))
142 pud
= pud_offset(pgd
, addr
);
143 if (!pud_present(*pud
))
146 pmd
= pmd_offset(pud
, addr
);
147 if (!pmd_present(*pmd
))
150 ptep
= pte_offset_map(pmd
, addr
);
152 if (!is_swap_pte(*ptep
)) {
157 ptl
= pte_lockptr(mm
, pmd
);
160 if (!is_swap_pte(pte
))
163 entry
= pte_to_swp_entry(pte
);
165 if (!is_migration_entry(entry
) || migration_entry_to_page(entry
) != old
)
169 * Yes, ignore the return value from a GFP_ATOMIC mem_cgroup_charge.
170 * Failure is not an option here: we're now expected to remove every
171 * migration pte, and will cause crashes otherwise. Normally this
172 * is not an issue: mem_cgroup_prepare_migration bumped up the old
173 * page_cgroup count for safety, that's now attached to the new page,
174 * so this charge should just be another incrementation of the count,
175 * to keep in balance with rmap.c's mem_cgroup_uncharging. But if
176 * there's been a force_empty, those reference counts may no longer
177 * be reliable, and this charge can actually fail: oh well, we don't
178 * make the situation any worse by proceeding as if it had succeeded.
180 mem_cgroup_charge(new, mm
, GFP_ATOMIC
);
183 pte
= pte_mkold(mk_pte(new, vma
->vm_page_prot
));
184 if (is_write_migration_entry(entry
))
185 pte
= pte_mkwrite(pte
);
186 flush_cache_page(vma
, addr
, pte_pfn(pte
));
187 set_pte_at(mm
, addr
, ptep
, pte
);
190 page_add_anon_rmap(new, vma
, addr
);
192 page_add_file_rmap(new);
194 /* No need to invalidate - it was non-present before */
195 update_mmu_cache(vma
, addr
, pte
);
198 pte_unmap_unlock(ptep
, ptl
);
202 * Note that remove_file_migration_ptes will only work on regular mappings,
203 * Nonlinear mappings do not use migration entries.
205 static void remove_file_migration_ptes(struct page
*old
, struct page
*new)
207 struct vm_area_struct
*vma
;
208 struct address_space
*mapping
= page_mapping(new);
209 struct prio_tree_iter iter
;
210 pgoff_t pgoff
= new->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
215 spin_lock(&mapping
->i_mmap_lock
);
217 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
)
218 remove_migration_pte(vma
, old
, new);
220 spin_unlock(&mapping
->i_mmap_lock
);
224 * Must hold mmap_sem lock on at least one of the vmas containing
225 * the page so that the anon_vma cannot vanish.
227 static void remove_anon_migration_ptes(struct page
*old
, struct page
*new)
229 struct anon_vma
*anon_vma
;
230 struct vm_area_struct
*vma
;
231 unsigned long mapping
;
233 mapping
= (unsigned long)new->mapping
;
235 if (!mapping
|| (mapping
& PAGE_MAPPING_ANON
) == 0)
239 * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
241 anon_vma
= (struct anon_vma
*) (mapping
- PAGE_MAPPING_ANON
);
242 spin_lock(&anon_vma
->lock
);
244 list_for_each_entry(vma
, &anon_vma
->head
, anon_vma_node
)
245 remove_migration_pte(vma
, old
, new);
247 spin_unlock(&anon_vma
->lock
);
251 * Get rid of all migration entries and replace them by
252 * references to the indicated page.
254 static void remove_migration_ptes(struct page
*old
, struct page
*new)
257 remove_anon_migration_ptes(old
, new);
259 remove_file_migration_ptes(old
, new);
263 * Something used the pte of a page under migration. We need to
264 * get to the page and wait until migration is finished.
265 * When we return from this function the fault will be retried.
267 * This function is called from do_swap_page().
269 void migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
,
270 unsigned long address
)
277 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
279 if (!is_swap_pte(pte
))
282 entry
= pte_to_swp_entry(pte
);
283 if (!is_migration_entry(entry
))
286 page
= migration_entry_to_page(entry
);
289 pte_unmap_unlock(ptep
, ptl
);
290 wait_on_page_locked(page
);
294 pte_unmap_unlock(ptep
, ptl
);
298 * Replace the page in the mapping.
300 * The number of remaining references must be:
301 * 1 for anonymous pages without a mapping
302 * 2 for pages with a mapping
303 * 3 for pages with a mapping and PagePrivate set.
305 static int migrate_page_move_mapping(struct address_space
*mapping
,
306 struct page
*newpage
, struct page
*page
)
311 /* Anonymous page without mapping */
312 if (page_count(page
) != 1)
317 write_lock_irq(&mapping
->tree_lock
);
319 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
322 if (page_count(page
) != 2 + !!PagePrivate(page
) ||
323 (struct page
*)radix_tree_deref_slot(pslot
) != page
) {
324 write_unlock_irq(&mapping
->tree_lock
);
329 * Now we know that no one else is looking at the page.
331 get_page(newpage
); /* add cache reference */
333 if (PageSwapCache(page
)) {
334 SetPageSwapCache(newpage
);
335 set_page_private(newpage
, page_private(page
));
339 radix_tree_replace_slot(pslot
, newpage
);
342 * Drop cache reference from old page.
343 * We know this isn't the last reference.
348 * If moved to a different zone then also account
349 * the page for that zone. Other VM counters will be
350 * taken care of when we establish references to the
351 * new page and drop references to the old page.
353 * Note that anonymous pages are accounted for
354 * via NR_FILE_PAGES and NR_ANON_PAGES if they
355 * are mapped to swap space.
357 __dec_zone_page_state(page
, NR_FILE_PAGES
);
358 __inc_zone_page_state(newpage
, NR_FILE_PAGES
);
360 write_unlock_irq(&mapping
->tree_lock
);
361 if (!PageSwapCache(newpage
)) {
362 mem_cgroup_uncharge_cache_page(page
);
369 * Copy the page to its new location
371 static void migrate_page_copy(struct page
*newpage
, struct page
*page
)
373 copy_highpage(newpage
, page
);
376 SetPageError(newpage
);
377 if (PageReferenced(page
))
378 SetPageReferenced(newpage
);
379 if (PageUptodate(page
))
380 SetPageUptodate(newpage
);
381 if (PageActive(page
))
382 SetPageActive(newpage
);
383 if (PageChecked(page
))
384 SetPageChecked(newpage
);
385 if (PageMappedToDisk(page
))
386 SetPageMappedToDisk(newpage
);
388 if (PageDirty(page
)) {
389 clear_page_dirty_for_io(page
);
391 * Want to mark the page and the radix tree as dirty, and
392 * redo the accounting that clear_page_dirty_for_io undid,
393 * but we can't use set_page_dirty because that function
394 * is actually a signal that all of the page has become dirty.
395 * Wheras only part of our page may be dirty.
397 __set_page_dirty_nobuffers(newpage
);
401 ClearPageSwapCache(page
);
403 ClearPageActive(page
);
404 ClearPagePrivate(page
);
405 set_page_private(page
, 0);
406 page
->mapping
= NULL
;
409 * If any waiters have accumulated on the new page then
412 if (PageWriteback(newpage
))
413 end_page_writeback(newpage
);
416 /************************************************************
417 * Migration functions
418 ***********************************************************/
420 /* Always fail migration. Used for mappings that are not movable */
421 int fail_migrate_page(struct address_space
*mapping
,
422 struct page
*newpage
, struct page
*page
)
426 EXPORT_SYMBOL(fail_migrate_page
);
429 * Common logic to directly migrate a single page suitable for
430 * pages that do not use PagePrivate.
432 * Pages are locked upon entry and exit.
434 int migrate_page(struct address_space
*mapping
,
435 struct page
*newpage
, struct page
*page
)
439 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
441 rc
= migrate_page_move_mapping(mapping
, newpage
, page
);
446 migrate_page_copy(newpage
, page
);
449 EXPORT_SYMBOL(migrate_page
);
453 * Migration function for pages with buffers. This function can only be used
454 * if the underlying filesystem guarantees that no other references to "page"
457 int buffer_migrate_page(struct address_space
*mapping
,
458 struct page
*newpage
, struct page
*page
)
460 struct buffer_head
*bh
, *head
;
463 if (!page_has_buffers(page
))
464 return migrate_page(mapping
, newpage
, page
);
466 head
= page_buffers(page
);
468 rc
= migrate_page_move_mapping(mapping
, newpage
, page
);
477 bh
= bh
->b_this_page
;
479 } while (bh
!= head
);
481 ClearPagePrivate(page
);
482 set_page_private(newpage
, page_private(page
));
483 set_page_private(page
, 0);
489 set_bh_page(bh
, newpage
, bh_offset(bh
));
490 bh
= bh
->b_this_page
;
492 } while (bh
!= head
);
494 SetPagePrivate(newpage
);
496 migrate_page_copy(newpage
, page
);
502 bh
= bh
->b_this_page
;
504 } while (bh
!= head
);
508 EXPORT_SYMBOL(buffer_migrate_page
);
512 * Writeback a page to clean the dirty state
514 static int writeout(struct address_space
*mapping
, struct page
*page
)
516 struct writeback_control wbc
= {
517 .sync_mode
= WB_SYNC_NONE
,
520 .range_end
= LLONG_MAX
,
526 if (!mapping
->a_ops
->writepage
)
527 /* No write method for the address space */
530 if (!clear_page_dirty_for_io(page
))
531 /* Someone else already triggered a write */
535 * A dirty page may imply that the underlying filesystem has
536 * the page on some queue. So the page must be clean for
537 * migration. Writeout may mean we loose the lock and the
538 * page state is no longer what we checked for earlier.
539 * At this point we know that the migration attempt cannot
542 remove_migration_ptes(page
, page
);
544 rc
= mapping
->a_ops
->writepage(page
, &wbc
);
546 /* I/O Error writing */
549 if (rc
!= AOP_WRITEPAGE_ACTIVATE
)
550 /* unlocked. Relock */
557 * Default handling if a filesystem does not provide a migration function.
559 static int fallback_migrate_page(struct address_space
*mapping
,
560 struct page
*newpage
, struct page
*page
)
563 return writeout(mapping
, page
);
566 * Buffers may be managed in a filesystem specific way.
567 * We must have no buffers or drop them.
569 if (PagePrivate(page
) &&
570 !try_to_release_page(page
, GFP_KERNEL
))
573 return migrate_page(mapping
, newpage
, page
);
577 * Move a page to a newly allocated page
578 * The page is locked and all ptes have been successfully removed.
580 * The new page will have replaced the old page if this function
583 static int move_to_new_page(struct page
*newpage
, struct page
*page
)
585 struct address_space
*mapping
;
589 * Block others from accessing the page when we get around to
590 * establishing additional references. We are the only one
591 * holding a reference to the new page at this point.
593 if (TestSetPageLocked(newpage
))
596 /* Prepare mapping for the new page.*/
597 newpage
->index
= page
->index
;
598 newpage
->mapping
= page
->mapping
;
600 mapping
= page_mapping(page
);
602 rc
= migrate_page(mapping
, newpage
, page
);
603 else if (mapping
->a_ops
->migratepage
)
605 * Most pages have a mapping and most filesystems
606 * should provide a migration function. Anonymous
607 * pages are part of swap space which also has its
608 * own migration function. This is the most common
609 * path for page migration.
611 rc
= mapping
->a_ops
->migratepage(mapping
,
614 rc
= fallback_migrate_page(mapping
, newpage
, page
);
617 remove_migration_ptes(page
, newpage
);
619 newpage
->mapping
= NULL
;
621 unlock_page(newpage
);
627 * Obtain the lock on page, remove all ptes and migrate the page
628 * to the newly allocated page in newpage.
630 static int unmap_and_move(new_page_t get_new_page
, unsigned long private,
631 struct page
*page
, int force
)
635 struct page
*newpage
= get_new_page(page
, private, &result
);
642 if (page_count(page
) == 1)
643 /* page was freed from under us. So we are done. */
646 charge
= mem_cgroup_prepare_migration(page
, newpage
);
647 if (charge
== -ENOMEM
) {
651 /* prepare cgroup just returns 0 or -ENOMEM */
655 if (TestSetPageLocked(page
)) {
661 if (PageWriteback(page
)) {
664 wait_on_page_writeback(page
);
667 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
668 * we cannot notice that anon_vma is freed while we migrates a page.
669 * This rcu_read_lock() delays freeing anon_vma pointer until the end
670 * of migration. File cache pages are no problem because of page_lock()
671 * File Caches may use write_page() or lock_page() in migration, then,
672 * just care Anon page here.
674 if (PageAnon(page
)) {
680 * Corner case handling:
681 * 1. When a new swap-cache page is read into, it is added to the LRU
682 * and treated as swapcache but it has no rmap yet.
683 * Calling try_to_unmap() against a page->mapping==NULL page will
684 * trigger a BUG. So handle it here.
685 * 2. An orphaned page (see truncate_complete_page) might have
686 * fs-private metadata. The page can be picked up due to memory
687 * offlining. Everywhere else except page reclaim, the page is
688 * invisible to the vm, so the page can not be migrated. So try to
689 * free the metadata, so the page can be freed.
691 if (!page
->mapping
) {
692 if (!PageAnon(page
) && PagePrivate(page
)) {
694 * Go direct to try_to_free_buffers() here because
695 * a) that's what try_to_release_page() would do anyway
696 * b) we may be under rcu_read_lock() here, so we can't
697 * use GFP_KERNEL which is what try_to_release_page()
698 * needs to be effective.
700 try_to_free_buffers(page
);
705 /* Establish migration ptes or remove ptes */
706 try_to_unmap(page
, 1);
708 if (!page_mapped(page
))
709 rc
= move_to_new_page(newpage
, page
);
712 remove_migration_ptes(page
, page
);
723 * A page that has been migrated has all references
724 * removed and will be freed. A page that has not been
725 * migrated will have kepts its references and be
728 list_del(&page
->lru
);
734 mem_cgroup_end_migration(newpage
);
736 * Move the new page to the LRU. If migration was not successful
737 * then this will free the page.
739 move_to_lru(newpage
);
744 *result
= page_to_nid(newpage
);
752 * The function takes one list of pages to migrate and a function
753 * that determines from the page to be migrated and the private data
754 * the target of the move and allocates the page.
756 * The function returns after 10 attempts or if no pages
757 * are movable anymore because to has become empty
758 * or no retryable pages exist anymore. All pages will be
759 * returned to the LRU or freed.
761 * Return: Number of pages not migrated or error code.
763 int migrate_pages(struct list_head
*from
,
764 new_page_t get_new_page
, unsigned long private)
771 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
775 current
->flags
|= PF_SWAPWRITE
;
777 for(pass
= 0; pass
< 10 && retry
; pass
++) {
780 list_for_each_entry_safe(page
, page2
, from
, lru
) {
783 rc
= unmap_and_move(get_new_page
, private,
795 /* Permanent failure */
804 current
->flags
&= ~PF_SWAPWRITE
;
806 putback_lru_pages(from
);
811 return nr_failed
+ retry
;
816 * Move a list of individual pages
818 struct page_to_node
{
825 static struct page
*new_page_node(struct page
*p
, unsigned long private,
828 struct page_to_node
*pm
= (struct page_to_node
*)private;
830 while (pm
->node
!= MAX_NUMNODES
&& pm
->page
!= p
)
833 if (pm
->node
== MAX_NUMNODES
)
836 *result
= &pm
->status
;
838 return alloc_pages_node(pm
->node
,
839 GFP_HIGHUSER_MOVABLE
| GFP_THISNODE
, 0);
843 * Move a set of pages as indicated in the pm array. The addr
844 * field must be set to the virtual address of the page to be moved
845 * and the node number must contain a valid target node.
847 static int do_move_pages(struct mm_struct
*mm
, struct page_to_node
*pm
,
851 struct page_to_node
*pp
;
854 down_read(&mm
->mmap_sem
);
857 * Build a list of pages to migrate
860 for (pp
= pm
; pp
->node
!= MAX_NUMNODES
; pp
++) {
861 struct vm_area_struct
*vma
;
865 * A valid page pointer that will not match any of the
866 * pages that will be moved.
868 pp
->page
= ZERO_PAGE(0);
871 vma
= find_vma(mm
, pp
->addr
);
872 if (!vma
|| !vma_migratable(vma
))
875 page
= follow_page(vma
, pp
->addr
, FOLL_GET
);
885 if (PageReserved(page
)) /* Check for zero page */
889 err
= page_to_nid(page
);
893 * Node already in the right place
898 if (page_mapcount(page
) > 1 &&
902 err
= isolate_lru_page(page
, &pagelist
);
905 * Either remove the duplicate refcount from
906 * isolate_lru_page() or drop the page ref if it was
914 if (!list_empty(&pagelist
))
915 err
= migrate_pages(&pagelist
, new_page_node
,
920 up_read(&mm
->mmap_sem
);
925 * Determine the nodes of a list of pages. The addr in the pm array
926 * must have been set to the virtual address of which we want to determine
929 static int do_pages_stat(struct mm_struct
*mm
, struct page_to_node
*pm
)
931 down_read(&mm
->mmap_sem
);
933 for ( ; pm
->node
!= MAX_NUMNODES
; pm
++) {
934 struct vm_area_struct
*vma
;
939 vma
= find_vma(mm
, pm
->addr
);
943 page
= follow_page(vma
, pm
->addr
, 0);
950 /* Use PageReserved to check for zero page */
951 if (!page
|| PageReserved(page
))
954 err
= page_to_nid(page
);
959 up_read(&mm
->mmap_sem
);
964 * Move a list of pages in the address space of the currently executing
967 asmlinkage
long sys_move_pages(pid_t pid
, unsigned long nr_pages
,
968 const void __user
* __user
*pages
,
969 const int __user
*nodes
,
970 int __user
*status
, int flags
)
974 struct task_struct
*task
;
975 nodemask_t task_nodes
;
976 struct mm_struct
*mm
;
977 struct page_to_node
*pm
= NULL
;
980 if (flags
& ~(MPOL_MF_MOVE
|MPOL_MF_MOVE_ALL
))
983 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
986 /* Find the mm_struct */
987 read_lock(&tasklist_lock
);
988 task
= pid
? find_task_by_vpid(pid
) : current
;
990 read_unlock(&tasklist_lock
);
993 mm
= get_task_mm(task
);
994 read_unlock(&tasklist_lock
);
1000 * Check if this process has the right to modify the specified
1001 * process. The right exists if the process has administrative
1002 * capabilities, superuser privileges or the same
1003 * userid as the target process.
1005 if ((current
->euid
!= task
->suid
) && (current
->euid
!= task
->uid
) &&
1006 (current
->uid
!= task
->suid
) && (current
->uid
!= task
->uid
) &&
1007 !capable(CAP_SYS_NICE
)) {
1012 err
= security_task_movememory(task
);
1017 task_nodes
= cpuset_mems_allowed(task
);
1019 /* Limit nr_pages so that the multiplication may not overflow */
1020 if (nr_pages
>= ULONG_MAX
/ sizeof(struct page_to_node
) - 1) {
1025 pm
= vmalloc((nr_pages
+ 1) * sizeof(struct page_to_node
));
1032 * Get parameters from user space and initialize the pm
1033 * array. Return various errors if the user did something wrong.
1035 for (i
= 0; i
< nr_pages
; i
++) {
1036 const void __user
*p
;
1039 if (get_user(p
, pages
+ i
))
1042 pm
[i
].addr
= (unsigned long)p
;
1046 if (get_user(node
, nodes
+ i
))
1050 if (!node_state(node
, N_HIGH_MEMORY
))
1054 if (!node_isset(node
, task_nodes
))
1059 pm
[i
].node
= 0; /* anything to not match MAX_NUMNODES */
1062 pm
[nr_pages
].node
= MAX_NUMNODES
;
1065 err
= do_move_pages(mm
, pm
, flags
& MPOL_MF_MOVE_ALL
);
1067 err
= do_pages_stat(mm
, pm
);
1070 /* Return status information */
1071 for (i
= 0; i
< nr_pages
; i
++)
1072 if (put_user(pm
[i
].status
, status
+ i
))
1083 * Call migration functions in the vma_ops that may prepare
1084 * memory in a vm for migration. migration functions may perform
1085 * the migration for vmas that do not have an underlying page struct.
1087 int migrate_vmas(struct mm_struct
*mm
, const nodemask_t
*to
,
1088 const nodemask_t
*from
, unsigned long flags
)
1090 struct vm_area_struct
*vma
;
1093 for(vma
= mm
->mmap
; vma
->vm_next
&& !err
; vma
= vma
->vm_next
) {
1094 if (vma
->vm_ops
&& vma
->vm_ops
->migrate
) {
1095 err
= vma
->vm_ops
->migrate(vma
, to
, from
, flags
);